From c6fd9a5c77fd1210ff65678c54ec1ba48e888318 Mon Sep 17 00:00:00 2001
From: Hermes Agent <hermes@obahan.xyz>
Date: Tue, 29 Sep 2026 19:54:56 +0000
Subject: [PATCH] Repair curly-quote JSON from LLM before parsing
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

halogen-qwen3.8-flash-next sometimes uses typographic quotes (“ ”) as JSON
string delimiters, which breaks strict json.loads. Add a scanner that
normalizes curly-delimited strings to straight quotes while preserving
curly quotes used as content inside straight-quoted strings.
---
 app/llm_client.py | 67 ++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 66 insertions(+), 1 deletion(-)

diff --git a/app/llm_client.py b/app/llm_client.py
index d79dcae..da44cb3 100644
--- a/app/llm_client.py
+++ b/app/llm_client.py
@@ -16,6 +16,71 @@ SYSTEM_PROMPT = (
 )
 
 
+def _try_parse(candidate: str):
+    try:
+        return json.loads(candidate)
+    except json.JSONDecodeError:
+        return None
+
+
+def _normalize_quotes(text: str) -> str:
+    """Rewrite LLM quote soup into valid JSON.
+
+    Handles models that use curly quotes (“ ”) as string DELIMITERS and also
+    as content inside straight-quoted strings. A scanner tracks which quote
+    character opened the current string so content quotes are preserved
+    (escaped) instead of breaking the structure.
+    """
+    out: list[str] = []
+    in_str: str | None = None  # '"' or '“'
+    i = 0
+    n = len(text)
+    while i < n:
+        c = text[i]
+        if in_str is None:
+            if c == '"':
+                in_str = '"'
+                out.append(c)
+            elif c == "\u201c":  # “ opens a string
+                in_str = "\u201c"
+                out.append('"')
+            else:
+                out.append(c)
+        else:
+            if c == "\\" and i + 1 < n:  # keep escape pairs intact
+                out.append(c)
+                out.append(text[i + 1])
+                i += 2
+                continue
+            if in_str == '"':
+                # Straight-delimited: curly quotes are just content.
+                out.append(c)
+                if c == '"':
+                    in_str = None
+            else:  # curly-delimited string
+                if c == "\u201d":  # ” closes it
+                    in_str = None
+                    out.append('"')
+                elif c == '"':  # raw straight quote inside -> escape
+                    out.append('\\"')
+                else:
+                    out.append(c)
+        i += 1
+    return "".join(out)
+
+
+def _repair_and_parse(text: str):
+    """Parse JSON, repairing common LLM quote mistakes if strict parse fails."""
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError:
+        pass
+    data = _try_parse(_normalize_quotes(text))
+    if data is not None:
+        return data
+    raise ValueError(f"Could not parse jokes JSON even after repair: {text[:200]!r}")
+
+
 def _extract_json_array(text: str) -> list[str]:
     """Pull a JSON array of strings out of a possibly messy LLM response."""
     text = text.strip()
@@ -28,7 +93,7 @@ def _extract_json_array(text: str) -> list[str]:
     end = text.rfind("]")
     if start == -1 or end == -1 or end <= start:
         raise ValueError(f"No JSON array found in LLM response: {text[:200]!r}")
-    data = json.loads(text[start : end + 1])
+    data = _repair_and_parse(text[start : end + 1])
     if not isinstance(data, list):
         raise ValueError("Parsed JSON is not a list")
     jokes = [str(j).strip() for j in data if str(j).strip()]